What NVIDIA announced
NVIDIA’s October 21 Agent Toolkit announcement includes Agent Knowledge Foundation and Kumo tools for predictions from structured business data. The developer overview separates Tabular for one table, Relational for connected tables, Forecast for future numeric signals, and Anomaly for unusual patterns. These are different jobs from writing a conversational answer. [1][2] This briefing covers an announcement made earlier today.
Some components predate this announcement. NVIDIA’s October 21 Forecast article describes it as formerly NV-Tesseract Forecasting, combining model predictions with optional retrieval of historical patterns. Keep the announcement date separate from the model version you evaluate. [3]
Check the component you will run
The official Forecast card lists Apache 2.0, while Relational weights use OpenMDW 1.1. Check the specific component’s terms. Forecast requires numeric time-series data; its card lists Linux and NVIDIA Ampere/Hopper compatibility. The overview links weights, separate runtimes, and NVIDIA NIM options. Confirm actual access, hardware, and compute costs for your chosen route. We have not installed or tested these tools. [2][3][4][5]
Our take
For experienced technicians, a useful first trial is a small demand forecast using data you have permission to process. Compare it with a simple baseline on later periods that were withheld from the model. Check errors across the forecast horizon, including unusual weeks and missing observations.
An explanation can help you investigate a prediction; it cannot establish that the prediction will hold. Keep a person responsible for decisions about stock, staffing, or customer outreach, and record which data and version produced the result.